Machine-learning augmented global analysis of urban emission trends, 2010-2022

Quantifying emission trends in urban systems globally is critical for guiding effective climate change mitigation solutions and informing major science assessments such as the reports by the Intergovernmental Panel on Climate Change. Currently available gridded emission datasets and self-reported inventories, however, are limited for analysing urban emissions, as downscaled products carry a significant national emission signal by construction, while self-reported estimates are biased by under-reporting incentives. Here, we develop a machine learning approach trained on a high quality bottom-up 1x1km 2 VULCAN inventory (RMSE = 39%, R 2 = 0.91) and apply it to similar cities globally. Comparing our estimates to ODIAC, we show that, on current trajectories, only 0-5% of global cities are on track to reach net zero and only 0-6% reduce emissions faster than their country. In 63-83% of cities, emissions are rising in lockstep with national trends. This is driven largely by Asian urban areas, where 88-95% of cities are growing emissions alongside their nations-though at a slower rate. Correlational analyses of the predictions indicate that cooling degree days is the most robust structural factor of urban-national emission divergence, though its direction reverses across regions with hotter cities lagging national trends in the Americas but outperforming in Europe, while population size shows no consistent association. These results demonstrate the value of machine learning for augmenting emission estimates at the subnational scale and highlight a stark gap between urban climate ambitions and current trajectories.